Home首页 · Glossary术语

Glossary术语

Each card takes one example from the labs and follows it from the table of numbers to the mistake a quick reading would make. The link at the bottom opens that page.每张卡片从实验里取一个例子，从那张数字表讲到读得太快会犯的错。底下的链接打开那一页。

Causal因果

Confounding混杂

A labour analyst has 300 workers and three columns: subsidy used or not, pay a year later, and household savings before the program. Workers with more savings both take the subsidy and earn more later. The analyst draws one line through pay and subsidy use, and the line slopes up. That slope is mixing the subsidy with the savings. The mix is confounding. If the subsidy itself adds nothing, the line can still slope up.一位劳工分析员有 300 名工人，三列数字：用没用补贴、一年后的月薪、参加前的家庭积蓄。积蓄更多的工人既去领补贴，后来月薪也更高。分析员在月薪和使用之间画一条线，线向上。这条线把补贴和积蓄搅在一起。这种搅在一起就是混杂。补贴本身没加一分钱，线仍可能向上。

Confounding lab混杂实验
Matching匹配

A department head has 80 students and three columns: joined the statistics workshop or not, a prior-math score, and the later methods grade. Students with higher prior math joined more often, so the grade gap can be that earlier score. Matching finds, for a joiner, a non-joiner with a close prior-math score, and compares those two grades. A student with no such partner is left out. The answer is then only about the pairs that could be made.一位系主任有 80 名学生，三列数字：参加统计工作坊与否、先前的数学分数、后来的方法课成绩。先前数学更高的人更常参加，所以成绩差距可能是那个旧分数。匹配为一名参加者找一名没参加、但先前数学接近的人，比较这两个人的成绩。找不到这样的人，就不进入比较。答案于是只关于能够配成的那些对。

Matching lab匹配实验
Caliper卡尺

The department head has to say how close two prior-math scores must be before the students count as a pair. That widest allowed gap is the caliper. At 2 points, many joiners find a non-joiner. At 0.3 points, most joiners find nobody, and they leave the comparison. The remaining pairs can look well matched. They are also a smaller group than the 80 students the department started with.系主任必须说，两个人的先前数学差多少以内，才算一对。这个允许的最大差距就是卡尺。差 2 分时，很多参加者找得到没参加的人。差 0.3 分时，大多数参加者找不到人，于是离开比较。剩下的配对可以看起来很接近。他们也比系里一开始的 80 人更少。

Matching lab匹配实验
Propensity倾向得分

The department head measured only prior math, and students with higher prior math were more likely to join the workshop. A propensity is one number that says how likely joining was, given the columns on the sheet. Here that number lines the 80 students up in the same order as prior math itself, because prior math is the only column. Matching on the propensity and matching on prior math then pick the same pairs.系主任只量了先前的数学，先前数学更高的学生更可能参加工作坊。倾向得分是一个数字，按表上的列，说出参加有多可能。这里这个数字把 80 名学生排成的顺序，和按先前数学排是一样的，因为先前数学是唯一的一列。按倾向得分匹配，和按先前数学匹配，会选出同样的配对。

Matching lab匹配实验

Networks网络

Stochastic block model随机块模型

A student-union survey lists 20 Arts students and 20 Engineering students, then asks each pair whether they talked last month. A stochastic block model draws those yes-or-no answers from two chances: one chance inside a faculty, another chance across faculties. The faculty names are on the sheet before any answer is drawn. The blank between the circles and the squares is where the dots were placed so the lines can be read. It is not a finding that the faculties split.学生会的调查列出文学院 20 人、工程学院 20 人，然后问每一对上个月有没有说过话。随机块模型用两个机会来画出这些是否答案：学院内部一个机会，跨学院另一个机会。学院的名字在任何答案画出之前就在表上。圆和方之间的空白，是为了把线看清楚而把点放开的位置。它不是“两个学院已经分开”这个发现。

Block-model lab块模型实验
Bridge桥

Among the 40 students, a bridge is one recorded talk between an Arts student and an Engineering student. The clinic notice can pass only along recorded talks. If Engineering has many talks inside the faculty and none to Arts, the notice can reach all 20 Engineering students and stop. The empty gap drawn between the two groups is not what stops it. The missing talk is.在这 40 名学生里，桥是一条记下的、文学院学生和工程学院学生之间的谈话。诊所通知只能沿记下的谈话传。如果工程学院内部谈话很多、通向文学院的一条都没有，通知可以传到全部 20 名工程学院学生，然后停下。画在两组之间的空隙不是挡住它的东西。缺失的那条谈话才是。

Diffusion lab扩散实验
Diffusion扩散

One Engineering student posts the clinic notice. Each day it can cross a recorded talk, with a probability you set, and it does not expire. Diffusion is that day-by-day passing. If the only talks are inside Engineering, the notice fills those 20 students and Arts stays unmarked. A full Engineering group is not evidence that Arts is about to see the notice. Arts sees it only after a talk crosses.工程学院的一名学生发出诊所通知。每一天它都可以按你设定的概率跨过一条记下的谈话，而且不会过期。扩散就是这种一天一天的传递。如果谈话只在工程学院内部，通知会填满这 20 人，文学院仍然没有标记。工程学院传满了，并不能说明文学院马上会看到。文学院要等一条谈话跨过去，才会看到。

Diffusion lab扩散实验

Machine learning机器学习

Overfitting过拟合

A housing analyst fits a curve to this year’s trust scores in 24 districts, using the share of households in public rental housing. The curve can bend until it touches nearly every district. Next year’s districts, drawn the same way, can sit far from that curve. The curve remembered this year’s wobbles. Overfitting is that too-close fit. A smaller miss on these 24 districts is not a reason to send next year’s outreach budget where the curve points.一位房屋分析员用公屋住户占比，去拟合今年 24 个区的信任分数。曲线可以一直弯，直到几乎碰到每一个区。按同样方式抽出的明年的区，可以离这条曲线很远。曲线记住的是今年的晃动。过拟合就是这种贴得太紧。这 24 个区上的误差更小，不是把明年的宣传预算投到曲线所指之处的理由。

Overfitting lab过拟合实验
Training and validation训练与验证

Training is the 24 districts the curve was drawn to touch. Validation is a fresh batch of districts from the same housing survey, looked at while the analyst is still choosing how flexible the curve should be. The curve can sit close to the 24 and far from the fresh batch. Because the analyst watched the fresh batch while choosing, that batch is still part of the choice. A last check would be districts nobody looked at until the curve was already fixed.训练是曲线被画去贴住的那 24 个区。验证是同一份房屋调查里新的一批区，分析员还在选曲线该多灵活时就拿来看。曲线可以贴近这 24 个区，却远离新的一批。因为挑选的时候新的一批是看得见的，它仍是挑选的一部分。最后的核对，是曲线已经定了以后才去看的那些区。

Overfitting lab过拟合实验
Mean squared error均方误差

For each of the 24 districts the curve predicts a trust score, and the prediction misses by some amount. Mean squared error averages those misses, and a miss of 4 counts more than four misses of 1. The number on the page is that average. A smaller number means the curve sat closer to this batch. It does not say the curve is the right one to use on next year’s districts.曲线为 24 个区里的每一个预测一个信任分数，预测都会偏一点。均方误差把这些偏差平均起来，偏了 4 的一次比偏了 1 的四次更重。页面上的数字就是这个平均。数字更小，只表示曲线更贴近这一批。它不说明这条曲线该用在明年的区上。

Overfitting lab过拟合实验
Ridge岭回归

The housing analyst keeps a degree-10 curve on the 24 districts, so every bend is still available. Ridge shortens every slope toward zero and almost never deletes one. The curve then moves less when the analyst draws another 24 districts. The cost is that a real bend, such as trust falling only in the districts with the highest public-rental share, can be shortened away too.房屋分析员在这 24 个区上保留一条 10 次的曲线，所以每一个弯都还在。岭回归把每个斜率往零缩短，几乎不删掉任何一个。分析员再抽 24 个区时，曲线就少动一些。代价是真正的弯，比如信任只在公屋占比最高的区下降，也可能被一起缩短掉。

Regularization lab正则化实验
Lasso套索

On the same degree-10 curve, lasso can set a slope to exactly zero, so its bar disappears from the chart. The analyst’s one-page note then names fewer housing features. A missing bar means this penalty dropped that feature on these 24 districts. Raise the penalty or draw the next year’s districts, and the bar can return. A blank bar is not a finding that the feature never mattered for trust.在同一条 10 次曲线上，套索可以把一个斜率正好设成零，于是它的条从图上消失。分析员的一页说明就少点一个房屋特征。缺掉的一条，是这个惩罚在这 24 个区上丢掉了那个特征。把惩罚调高，或换明年的区，这条可以回来。空白的一条，不是“这个特征对信任从来无关”的发现。

Regularization lab, lasso正则化实验，套索
Penalty λ惩罚 λ

λ is how hard the penalty pulls the slopes of the 24-district curve toward zero. At 0, every slope can chase the districts. At a high value, the curve is nearly flat and ignores the public-rental share. The same λ is not the same pull if the analyst rescales that share from a fraction to a percent. The number belongs to this scaling of these columns.λ 是惩罚把这 24 个区的曲线的斜率往零拉得多用力。为 0 时，每个斜率都可以去追这些区。数值很高时，曲线几乎变平，公屋占比被忽略。如果分析员把这个占比从小数改成百分数，同一个 λ 就不是同样的拉力。这个数字属于这些列的这种缩放。

Regularization lab正则化实验
Pooled slope汇总斜率

A registrar has 12 students in each of four faculties, with workshop hours and a course grade. Engineering’s line can be steep and Arts’ line flat. A pooled slope ignores the faculty name and draws one line through all 48. That line often lies between the four and matches none of them. Reporting it as “the effect of the workshop” names a student who is not in any of the four faculties.一位教务主任在四个学院各有 12 名学生，列着工作坊时数和课程成绩。工程学院的线可以很陡，文学院的线很平。汇总斜率不管学院的名字，给全部 48 人画一条线。这条线常常落在四条线中间，哪一条都不像。把它报告成“工作坊的效应”，说的是一个不在这四个学院里的学生。

Group-slopes lab分组斜率实验
Prior先验

A district councillor already has 24 consultation forms from last year on a pedestrian scheme, and about 30 percent of them supported it. That pile, and how much it should count, is the prior. It is on the desk before this week’s 12 interviews arrive. If the councillor treats the 24 forms as a heavy pile, 8 new supporters out of 12 barely move the share. If the pile is treated as light, those 12 interviews move it at once.一位区议员已经有去年 24 份关于步行街方案的咨询表，其中大约三成表示支持。这一叠，以及它该算多重，就是先验。它在本周 12 份访谈到来之前就在桌上。如果议员把这 24 份算得很重，12 人里新的 8 个支持者几乎推不动这个比例。如果这一叠算得很轻，这 12 份访谈立刻就把它推走。

Updating lab更新实验
Posterior后验

The 12 interviews arrive: 8 support the scheme, 4 do not. The posterior is last year’s 24 forms after those 12 have been allowed to move the share. With a heavy pile of old forms, Monday’s number stays near 30 percent. With a light pile, or with 60 interviews instead of 12, Monday’s number sits near 8 out of 12. A different set of last year’s forms would have landed somewhere else. The solid curve is that updated share.12 份访谈到了：8 人支持方案，4 人不支持。后验是去年的 24 份表在这 12 份被允许推动比例之后的位置。旧表很重时，周一写下的数字留在三成附近。旧表很轻，或访谈是 60 份而不是 12 份时，周一的数字靠近 12 人中的 8 人。换一批去年的表，它会落在别处。实线就是这个更新后的比例。

Updating lab更新实验

Text as data文本作为数据

Bag of words词袋

A press officer has two headlines: “The new training program raised wages for the workers in the plant” and the same line with “lowered” instead of “raised.” A bag of words counts each word and throws away the order. The two headlines then share almost every count, and the one verb that flips the claim is a small part of the score. A high score can be written up as “the coverage agrees.” The counts never saw which verb came first.一位新闻官有两则标题：“The new training program raised wages for the workers in the plant”，以及把 raised 换成 lowered 的同一句。词袋数每个词，并丢掉顺序。两则标题于是几乎共享每一个计数，把主张说反的那个动词只占分数的一小部分。高分可以被写成“报道口径一致”。计数从没看见哪个动词在前。

Bag-of-words lab词袋实验
Document–term matrix文档—词矩阵

Put the two headlines in a table. Each row is one headline. Each column is one word, such as training, wages, raised, or lowered. The cell is how many times that headline used that word. The two rows look alike because they share training, wages, and workers. The table does not record that one headline says pay went up and the other says pay went down. A document–term matrix is that table.把两则标题放进一张表。每一行是一则标题。每一列是一个词，比如 training、wages、raised 或 lowered。格子是这则标题用了这个词几次。两行看起来像，因为它们都有 training、wages 和 workers。表没有记下：一则说工资涨了，另一则说工资降了。文档—词矩阵就是这张表。

Bag-of-words lab词袋实验
Cosine similarity余弦相似度

Cosine is the score the press officer gets from the two rows of word counts. It asks whether the rows point the same way, and it ignores which headline is longer. The raised/lowered pair scores high because both rows are full of the, training, program, wages, and workers. The verb that flips the claim barely moves the number. A high cosine is a resemblance of those counts. It is not a reading of whether pay rose.余弦是新闻官从两行词数得到的分数。它问两行是否朝向同一个方向，并且不管哪则标题更长。raised 和 lowered 那一对分数高，因为两行都堆满了 the、training、program、wages 和 workers。把主张说反的那个动词几乎挪不动这个数。余弦高，是这些计数像。不是读出了工资到底涨没涨。

Bag-of-words lab词袋实验
Topic mixture主题混合

The press officer has nine clippings and has to choose how many folders to use before Monday. A topic mixture would let one clipping be part wage story and part model story at the same time. This page does not do that. It gives each clipping one color. With one folder, wage words, tie words, and model words share a list. With three folders, the colors can match the three lists the clippings were written as. The folder count was chosen. It was not found in the clippings.新闻官有九行剪报，周一前必须决定用几个文件夹。主题混合会让一行剪报同时是工资故事的一部分，又是模型故事的一部分。这一页不这么做。它给每行一种颜色。一个文件夹时，工资用词、关系用词和模型用词共用一份词表。三个文件夹时，颜色可以和这些剪报被写成的三份清单对上。文件夹的个数是选定的。它不是从剪报里找出来的。

Topic-count lab主题个数实验
Embedding嵌入

An embedding gives each word a place, so the press officer can score headlines that do not share spelling. On this page the places were typed: raised sits opposite lowered, and pay sits near earnings. A headline’s mark is the average of its words’ places. The raised/lowered pair then sits far apart on the map even though the word counts were almost the same. A trained embedding would learn the places from many headlines. These places were typed, so the map already separates the verb that flips the claim.嵌入给每个词一个位置，于是新闻官可以给拼写不同的标题打分。这一页的位置是敲进去的：raised 在 lowered 对面，pay 靠近 earnings。一则标题的标记是它那些词的位置的平均。raised 和 lowered 那一对于是在图上离得很远，尽管词数几乎一样。训练出来的嵌入会从很多标题里学习这些位置。这些位置是敲进去的，所以地图已经把那个说反主张的动词分开了。

Word-map lab词图实验

Glossary术语

Causal因果

Confounding混杂

Matching匹配

Caliper卡尺

Propensity倾向得分

Networks网络

Stochastic block model随机块模型

Bridge桥

Diffusion扩散

Machine learning机器学习

Overfitting过拟合

Training and validation训练与验证

Mean squared error均方误差

Ridge岭回归

Lasso套索

Penalty λ惩罚 λ

Pooled slope汇总斜率

Prior先验

Posterior后验

Text as data文本作为数据

Bag of words词袋

Document–term matrix文档—词矩阵

Cosine similarity余弦相似度

Topic mixture主题混合

Embedding嵌入